Search

Land cover mapping is essential to understanding global land-use patterns and studying biodiversity composition and the functioning of eco-systems. The introduction of remote sensing technologies and artificial intelligence models made it possible to base land cover mapping on satellite imagery in order to monitor changes, assess ecosystem health, support conservation efforts, and reduce monitoring time. However, significant challenges remain in managing large, complex satellite imagery datasets, acquiring specialized datasets due to high costs and labor intensity, including a lack of comparative studies for the selection of optimal deep learning models. No less important is the scarcity of aerial datasets specifically tailored for agricultural areas. This study addresses these gaps by presenting a methodology for semantic segmentation of land covers in agricultural areas using satellite images and deep learning models with pre-trained backbones. We introduce an efficient methodology for preparing semantic segmentation datasets and contribute the “Land Cover Aerial Imagery” (LICAI) dataset for semantic segmentation. The study focuses on the Franciacorta area, Lombardy Region, leveraging the rich diversity of the dataset to effectively train and evaluate the models. We conducted a comparative study, using cutting-edge deep-learning-based segmentation models (U-Net, SegNet, DeepLabV3) with various pre-trained backbones (ResNet, Inception, DenseNet, EfficientNet) on our dataset acquired from Google Earth Pro. Through meticulous data acquisition, preprocessing, model selection, and evaluation, we demonstrate the effectiveness of these techniques in accurately identifying land cover classes. Integrating pre-trained feature extraction networks significantly improves performance across various metrics. Additionally, addressing challenges such as data availability, computational resources, and model interpretability is essential for advancing the field of remote sensing, in support of biodiversity conservation and the provision of ecosystem services and sustainable agriculture.

Radiology reports are a valuable source of textual information used to improve clinical care and support research. In recent years, deep learning techniques have been shown to be effective in classifying radiology reports. This article investigates the use of deep learning techniques with attention mechanisms to achieve better performance in the classification of radiology reports. We focus on various Natural Language Processing approaches, such as LSTM with Attention, BERT, and GPT-4, evaluated on a chest tomography report dataset regarding neoplastic diseases collected from an Italian hospital. In particular, we compare the results with a previous machine learning system, showing that models based on attention mechanisms can achieve higher performance. The Attention Mechanism allows us to identify the most relevant bits of text used by the model to make its predictions. We show that our model achieves state-of-the-art results on the hierarchical classification of radiology reports. Moreover, we evaluate the performance of GPT-4 on the classification of these reports in a zero-shot setup through prompt engineering, showing interesting results even with a small context and a non-English language. Our findings suggest that deep learning techniques with attention mechanisms may be successful in the classification of radiology reports even in non-English languages for which it is not possible to leverage on large text corpus.

The analysis of the presence of bias, prejudices and unwanted discriminatory behavior in pre-trained neural language models (NLMs), considering the sensitivity of the topic and its public interest, should respect two main criteria: the intuition and the statistical rigor. To the state of the art, there are two main categories of approaches for analyzing bias: those based on the models’ textual output, and those based on the geometric space of the embedded representations calculated by the NLMs. While the first one is intuitive, this kind of analysis is often conducted on simple template sentences, which limit the overall validity of their conclusions in a real-world context. On the contrary, geometric methods are more rigorous but quite more complex to implement and understand for those who are non-experts in Natural Language Processing (NLP). In this paper, we propose a unique method for analyzing bias in pre-trained language models that combines these two aspects. Through a simple classification task, we verify whether the information contained in the embedded representation of words that describes a protected property (such as the religion) can be used to identify a stereotyped property (such as the criminal behavior), requiring only a minimal supervised dataset. We experimentally verify our approach, finding that four widespread Transformer-based models are affected by prejudices of gender, nationality, and religion.

The automatic identification of plant species using unmanned aerial vehicles (UAVs) is a valuable tool for ecological research. However, challenges such as reduced spatial resolution due to high-altitude operations, image degradation from camera optics and sensor limitations, and information loss caused by terrain shadows hinder the accurate classification of plant species from UAV imagery. This study addresses these issues by proposing a novel image preprocessing pipeline and evaluating its impact on model performance. Our approach improves image quality through a multi-step pipeline that includes Enhanced Super-Resolution Generative Adversarial Networks (ESRGAN) for resolution enhancement, Contrast-Limited Adaptive Histogram Equalization (CLAHE) for contrast improvement, and white balance adjustments for accurate color representation. These preprocessing steps ensure high-quality input data, leading to better model performance. For feature extraction and classification, we employ a pre-trained VGG-16 deep convolutional neural network, followed by machine learning classifiers, including Support Vector Machine (SVM), random forest (RF), and Extreme Gradient Boosting (XGBoost). This hybrid approach, combining deep learning for feature extraction with machine learning for classification, not only enhances classification accuracy but also reduces computational resource requirements compared to relying solely on deep learning models. Notably, the VGG-16 + SVM model achieved an outstanding accuracy of 97.88% on a dataset preprocessed with ESRGAN and white balance adjustments, with a precision of 97.9%, a recall of 97.8%, and an F1 score of 0.978. Through a comprehensive comparative study, we demonstrate that the proposed framework, utilizing VGG-16 for feature extraction, SVM for classification, and preprocessed images with ESRGAN and white balance adjustments, achieves superior performance in plant species identification from UAV imagery.

AI and Machine Learning (ML) offer powerful tools to support clinical decision making in emergency situations such as the COVID-19 pandemic. In this context, the application of ML requires to design predictive systems that have adequate accuracy and can effectively deal with issues concerning data quality, sensitive errors, uncertainty, and interpretability of the predictions. We present a methodology that deals with all these problems and a concrete study of its application to estimate the prognosis of hospitalised patients with COVID-19. In particular, we address the task of predicting the outcome (alive or deceased) of a patient at different times of her/his hospitalisation minimising false negatives (wrong survival predictions). The proposed methodology builds different optimised ML models to select those that perform the best to recognise, at different times of hospitalisation, patients who will have an unfavourable prognosis (decease). These models exploit a new algorithm, presented in the paper, that identifies an uncertainty threshold to rule out uncertain predictions with the purpose of making a ML model both more performing and more reliable. Moreover, we propose a general method for automatically extracting multi-variable prognostic rules from the available data. Such rules can provide possible new useful knowledge on the considered disease. We also show how they can be used effectively to explain the predictions made by the ML models. All proposed methods and techniques are experimentally evaluated in the context of our application task.

Biodiversity regulates agroecosystem processes, ensuring stability. Preserving and restoring biodiversity is vital for sustainable agricultural production. Species identification and classification in plant communities are key in biodiversity studies. Remote sensing supports species identification. However, accurately identifying plant species in heterogeneous plant areas presents challenges in dataset acquisition, preparation, and model selection for image classification. This study presents a method that combines object-based supervised machine learning for dataset preparation and a pre-trained transfer learning model (EfficientNetV2) for precise plant species classification in heterogeneous areas. The methodology is based on the multi-resolution segmentation of the UAV RGB orthophoto of the plant community into multiple canopy objects, and on the classification of the plants in the orthophoto using the K-nearest neighbor (KNN) supervised machine learning algorithm. Individual plant species canopies are extracted with the ArcGIS training dataset. A pre-trained transfer learning model is then applied for classification. Test results show that the EfficientNetV2 achieves an impressive 99% classification accuracy for seven plant species. A comparative study contrasts the EfficientNetV2 model with other widely used transfer learning models: ResNet50, Xception, DenseNet121, InceptionV3, and MobileNetV2.

Artificial Intelligence and Natural Language Processing techniques can have a very significant impact on the e-learning sector, with the introduction of chatbots, automatic correctors, or scoring systems. However, integrating such technologies into the business environment in an effective way is not a trivial operation, and it not only requires realising a model with good predictive performance, but also it requires the following: (i) a proper study of the task, (ii) a data collection process, (iii) a real-world evaluation of its utility. Moreover, it is also very important to build an entire IT infrastructure that connects the AI system with the company database, with the human employees, the users, etc. In this work, we present a real-world system, based on the state-of-the-art BERT model, which implements an automatic scoring system for open-ended questions written in Italian. More specifically, these questions pertain to the workplace safety courses which every worker must attend by law, often via e-learning platforms such as the one offered by Mega Italia Media. This article describes how our system has been designed, evaluated, and finally deployed for commercial use with complete integration with the other services provided by the company.

In recent years, many studies have been devoted to discovering the inner workings of Transformer-based models, such as BERT, for instance, attempting to identify what information is contained within them. However, little is known about how these models store this information in their millions of parameters and which parts of the architecture are the most important. In this work, we propose an approach to identify self-attention mechanisms, called heads, that contain semantic and real-world factual knowledge in BERT. Our approach includes a metric computed from attention weights and exploits a standard clustering algorithm for extracting the most relevant connections between tokens in a head. In our experimental analysis, we focus on how heads can connect synonyms, antonyms and several types of factual knowledge regarding subjects such as geography and medicine.

Single-task models (STMs) struggle to learn sophisticated representations from a finite set of annotated data. Multitask learning approaches overcome these constraints by simultaneously training various associated tasks, thereby learning generic representations among various tasks by sharing some layers of the neural network architecture. Because of this, multitask models (MTMs) have better generalization properties than those of single-task learning. Multitask model generalizations can be used to improve the results of other models. STMs can learn more sophisticated representations in the training phase by utilizing the extracted knowledge of an MTM through the knowledge distillation technique where one model supervises another model during training by using its learned generalizations. This paper proposes a knowledge distillation technique in which different MTMs are used as the teacher model to supervise different student models. Knowledge distillation is applied with different representations of the teacher model. We also investigated the effect of the conditional random field (CRF) and softmax function for the token-level knowledge distillation approach, and found that the softmax function leveraged the performance of the student model compared to CRF. The result analysis was also extended with statistical analysis by using the Friedman test.

Biomedical named entity recognition (BioNER) is a preliminary task for many other tasks, e.g., relation extraction and semantic search. Extracting the text of interest from biomedical documents becomes more demanding as the availability of online data is increasing. Deep learning models have been adopted for biomedical named entity recognition (BioNER) as deep learning has been found very successful in many other tasks. Nevertheless, the complex structure of biomedical text data is still a challenging aspect for deep learning models. Limited annotated biomedical text data make it more difficult to train deep learning models with millions of trainable parameters. The single-task model, which focuses on learning a specific task, has issues in learning complex feature representations from a limited quantity of annotated data. Moreover, manually constructing annotated data is a time-consuming job. It is, therefore, vital to exploit other efficient ways to train deep learning models on the available annotated data. This work enhances the performance of the BioNER task by taking advantage of various knowledge transfer techniques: multitask learning and transfer learning. This work presents two multitask models (MTMs), which learn shared features and task-specific features by implementing the shared and task-specific layers. In addition, the presented trained MTM is also fine-tuned for each specific dataset to tailor it from a general features representation to a specialized features representation. The presented empirical results and statistical analysis from this work illustrate that the proposed techniques enhance significantly the performance of the corresponding single-task model (STM).

During the COVID-19 pandemic, the corporate online training sector has increased exponentially and online course providers had to implement innovative solutions to be more efficient and provide a satisfactory service. This paper considers a real case study in implementing a chatbot, which answers frequently asked questions from learners on an Italian e-learning platform that provides workplace safety courses to several business customers. Having to respond quickly to the increase in the courses activated, the company decided to develop a chatbot using a cloud-based service currently available on the market. These services are based on Natural Language Understanding (NLU) engines, which deal with identifying information such as entities and intentions from the sentences provided as input. To integrate a chatbot in an e-learning platform, we studied the performance of the intent recognition task of the major NLU platforms available on the market with an in-depth comparison, using an Italian dataset provided by the owner of the e-learning platform. We focused on intent recognition, carried out several experiments and evaluated performance in terms of F-score, error rate, response time, and robustness of all the services selected. The chatbot is currently in production, therefore we present a description of the system implemented and its results on the original users’ requests.